diff --git a/python/paddle_serving_server/__init__.py b/python/paddle_serving_server/__init__.py
index 6da0e6597e69d605b7172139d32c00b77d538386..68a3b424cd9bda861ddf0b54ac42bf356c3de310 100644
--- a/python/paddle_serving_server/__init__.py
+++ b/python/paddle_serving_server/__init__.py
@@ -260,6 +260,7 @@ class Server(object):
         self._prepare_resource(workdir)
         self._prepare_engine(self.model_config_path, device)
         self._prepare_infer_service(port)
+        self.port = port
         self.workdir = workdir
 
         infer_service_fn = "{}/{}".format(workdir, self.infer_service_fn)
diff --git a/python/paddle_serving_server/web_service.py b/python/paddle_serving_server/web_service.py
index ca43426c2a82a0c8be296c8410361acbf498fc5c..31b786bcb53b5c6724549ac19f94463c01c1f680 100755
--- a/python/paddle_serving_server/web_service.py
+++ b/python/paddle_serving_server/web_service.py
@@ -18,6 +18,8 @@ from flask import Flask, request, abort
 from multiprocessing import Pool, Process
 from paddle_serving_server import OpMaker, OpSeqMaker, Server
 from paddle_serving_client import Client
+from contextlib import closing
+import socket
 
 
 class WebService(object):
@@ -41,19 +43,34 @@ class WebService(object):
         server.set_num_threads(16)
         server.load_model_config(self.model_config)
         server.prepare_server(
-            workdir=self.workdir, port=self.port + 1, device=self.device)
+            workdir=self.workdir, port=self.port_list[0], device=self.device)
         server.run_server()
 
+    def port_is_available(self, port):
+        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
+            sock.settimeout(2)
+            result = sock.connect_ex(('0.0.0.0', port))
+        if result != 0:
+            return True
+        else:
+            return False
+
     def prepare_server(self, workdir="", port=9393, device="cpu"):
         self.workdir = workdir
         self.port = port
         self.device = device
+        default_port = 6000
+        self.port_list = []
+        for i in range(1000):
+            if self.port_is_available(default_port + i):
+                self.port_list.append(default_port + i)
+                break
 
     def _launch_web_service(self):
         self.client_service = Client()
         self.client_service.load_client_config(
             "{}/serving_server_conf.prototxt".format(self.model_config))
-        self.client_service.connect(["0.0.0.0:{}".format(self.port + 1)])
+        self.client_service.connect(["0.0.0.0:{}".format(self.port_list[0])])
 
     def get_prediction(self, request):
         if not request.json:
diff --git a/python/paddle_serving_server_gpu/web_service.py b/python/paddle_serving_server_gpu/web_service.py
index b4c911b7d805dab376a9709ce1ca5d719ea49794..abec2b55f31fbd4a99b3f50aaaa1a4851de316fb 100644
--- a/python/paddle_serving_server_gpu/web_service.py
+++ b/python/paddle_serving_server_gpu/web_service.py
@@ -13,14 +13,15 @@
 # limitations under the License.
 
 from flask import Flask, request, abort
-from paddle_serving_server_gpu import OpMaker, OpSeqMaker, Server
-import paddle_serving_server_gpu as serving
+from contextlib import closing
 from multiprocessing import Pool, Process, Queue
 from paddle_serving_client import Client
+from paddle_serving_server_gpu import OpMaker, OpSeqMaker, Server
 from paddle_serving_server_gpu.serve import start_multi_card
-
+import socket
 import sys
 import numpy as np
+import paddle_serving_server_gpu as serving
 
 
 class WebService(object):
@@ -66,22 +67,39 @@ class WebService(object):
     def _launch_rpc_service(self, service_idx):
         self.rpc_service_list[service_idx].run_server()
 
+    def port_is_available(self, port):
+        with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
+            sock.settimeout(2)
+            result = sock.connect_ex(('0.0.0.0', port))
+        if result != 0:
+            return True
+        else:
+            return False
+
     def prepare_server(self, workdir="", port=9393, device="gpu", gpuid=0):
         self.workdir = workdir
         self.port = port
         self.device = device
         self.gpuid = gpuid
+        self.port_list = []
+        default_port = 6000
+        for i in range(1000):
+            if self.port_is_available(default_port + i):
+                self.port_list.append(default_port + i)
+            if len(self.port_list) > len(self.gpus):
+                break
+
         if len(self.gpus) == 0:
             # init cpu service
             self.rpc_service_list.append(
                 self.default_rpc_service(
-                    self.workdir, self.port + 1, -1, thread_num=10))
+                    self.workdir, self.port_list[0], -1, thread_num=10))
         else:
             for i, gpuid in enumerate(self.gpus):
                 self.rpc_service_list.append(
                     self.default_rpc_service(
                         "{}_{}".format(self.workdir, i),
-                        self.port + 1 + i,
+                        self.port_list[i],
                         gpuid,
                         thread_num=10))
 
@@ -93,9 +111,9 @@ class WebService(object):
         endpoints = ""
         if gpu_num > 0:
             for i in range(gpu_num):
-                endpoints += "127.0.0.1:{},".format(self.port + i + 1)
+                endpoints += "127.0.0.1:{},".format(self.port_list[i])
         else:
-            endpoints = "127.0.0.1:{}".format(self.port + 1)
+            endpoints = "127.0.0.1:{}".format(self.port_list[0])
         self.client.connect([endpoints])
 
     def get_prediction(self, request):